Blog
Product updates, engineering deep-dives, and lessons learned building AI-powered code reviews.

Someone asked whether code quality was improving. We had 400 review comments and no answer. So we built the thing that would have made the answer obvious.

The bot kept flagging something your team does on purpose. You'd dismiss the comment, and next PR it came back. Now you can just reply — and it remembers.

You open a PR on Monday, tag two reviewers, and wait. By Wednesday one has left 'LGTM' and the other hasn't looked. We built something so that never happens again.

"GitHub tells us, we analyze, we comment." That's the short answer. The interesting part is the ten steps in between — and why each one exists.

An early user told us they'd stopped reading the bot's comments. Not because the feedback was wrong — because it repeated itself every time they pushed. Here's how we fixed it.